well new_addr - original_section but yeah
#EvalynOS
rancid viper
hazy saddle
And then for static GS I can still do gs:whatever and have a fixed struct at the top of a section
rancid viper
you'd need to place the newly allocated sections after the kernel afaik
hazy saddle
I don’t think my allocator can handle that
rancid viper
it may be easier to just have macros
because you'd need to use __atomic* accessors or whatever anyway
to guarantee atomicity
hazy saddle
Yeah
Il have a macro that takes in the variable and reruns the pointer to it based on GS?
Or the data inside or etc
rotund furnace
And what I do since I shamelessly stole the macros directly :3
magic saffron
the compiler knows the offset of data from gs_base because it knows where it put the data in that section, similar to how we discussed it somewhere somewhere else (and how managarm does it). The difference here is that the compiler does it all for you. Its the same as it knowing how to access int x;, just using gs: instead of an implicit ds: (with assumed 0 base).
__seg_gs is convinient but the big downside is that pointers are only valid in the "gs address space", not the default one.
actually the real downside is its not compatible with c++ semantics, so its C only 😔
hazy saddle
i think ive manages to swap over CPU local
but this runs slow
nvm its just this laptop
the QEMU framebuffer is slow as balls
gentle quest
if u wanna see an example of automatic cpu locals with seg gs i do it like that in my kernel
so for me a cpu local load costs the same amount of operations as loading any variable
e.g.
mov eax,DWORD PTR gs:[rip+0x1da40] # ffffffff800280b0 <g_this_cpu_id>
the trick is what i load into GS_BASE is not actually a gs base, rather it's an addend that must be added to the .per_cpu section to make the address overflow into the dynamically allocated per-cpu section
and that addition is done by cpu transparently
also with this setup u get super early cpu locals by just setting gs base to 0
and if u end up in a single cpu environment u can just continue using that section
hazy saddle
the trick is what i load into GS_BASE is not actually a gs base, rather it's an ...
When you do this how do you get the thread block on syscalls? So you can do swap swapping etc
gentle quest
wdym?
hazy saddle
On syscall I have to swap stacks right
Before I had CPU local I stored the thread block directly under GS
But with this I need an extra pointer deref
Which needs another register
gentle quest
linux uses a field in the tss as scratch space for rsp
SYM_CODE_START(entry_SYSCALL_64)
UNWIND_HINT_ENTRY
ENDBR
swapgs
/* tss.sp2 is scratch space. */
movq %rsp, PER_CPU_VAR(cpu_tss_rw + TSS_sp2)
SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp
yep
rotund furnace
using rsp as the kpti scratch reg is insanely smart
hazy saddle
So the TSS is stored fully inside of cpu local on linux?
And then it just uses a spot inside that?
gentle quest
all of cpu locals are stored in cpu local
hazy saddle
gentle quest
u just duplicate the entire .per_cpu section for all cpus
hazy saddle
So I can technically just do what I was doing withhh
rotund furnace
wait linux dumps the tss inside the cpu local struct
gentle quest
also linux has dynamic per cpu variables
rotund furnace
fuck that's kinda smart
gentle quest
yeah
gentle quest
wait linux dumps the tss inside the cpu local struct
there isnt a cpu local struct
thats the thing
neon karma
also linux has dynamic per cpu variables
what does that mean?
hazy saddle
So my idea wasn’t even dumb and I redid all this for nothing
gentle quest
u have to understand its just a normal binary section
hazy saddle
neon karma
what could that be used for
gentle quest
anything that can be used per cpu, counters, queues, stuff like that
neon karma
sure but why not a global cpulocal variable
gentle quest
wdym?
neon karma
the thing I do now
gentle quest
if u wanna see examples https://elixir.bootlin.com/linux/v6.19.9/A/ident/alloc_percpu
gentle quest
sure but why not a global cpulocal variable
u mean as opposed to just a section?
hazy saddle
https://cdn.discordapp.com/attachments/1434603991957897258/1483237320978202925/i...
Is GSbase = top of CPUlocal data and then calculating the offset still considered atomic/one instruction?
gentle quest
because modules may introduce their own percpu variables
and those are loaded dynamically
gentle quest
also shoving unrelated variables into one struct doesnt really achieve anything
and a section just automatically groups unrelated variables
u can have a static per-cpu variable and it will work
without global visilibity
hazy saddle
Is GSbase = top of CPUlocal data and then calculating the offset still considere...
Or do I still need the fancy delta stuff
Though it is cheaper with that
gentle quest
Is GSbase = top of CPUlocal data and then calculating the offset still considere...
what
gsbase = pcpu_base - .per-cpu-section
hazy saddle
this_cpu_offset = g_per_cpu_base;
this_cpu_offset -= (ptr_t)SECTION_MARKER_BEGIN(PER_CPU_SECTION);
then its gs:varrible?
hazy saddle
whats the proper way to refer to that inside assembly?
that also lets me use clang address spaces right i think too?
gentle quest
%gs:var(%rip)
gentle quest
that also lets me use clang address spaces right i think too?
__seg_gs is actually an alias for address space
hazy saddle
?? i mean how do i know how to refer to that var from inside asm?
gentle quest
`%gs:var(%rip)`
^
hazy saddle
what wiuld var be in this case?
gentle quest
the per cpu variable u want to read
hazy saddle
but how do i know what that is in assembly?? where i cant refer to the C linker stuff?
hazy saddle
because if the varrible is defined in C how do i refer to it from assembly?
gentle quest
the workflow is
DEFINE_PER_CPU(current_thread);
in asm:
mov %gs:current_thread(%rip), rax
would the offset be like :0 if its at the very start of perCPU
gentle quest
no, it would be the offset from end_of_kernel_half - 2GiB + .per_cpu
not that u care anyway because the compiler puts the right number by itself
thats why this approach is so cool
hazy saddle
so how do i get that into assembly??? because i dont know what it is at all from there?
gentle quest
get what into assembly
hazy saddle
if i need to refer to CPU local things in asssembly
gentle quest
i literally showed so many examples 😭
gentle quest
how do u access normal variables from assembly
wanton grove
`gsbase = pcpu_base - .per-cpu-section`
why people need per cpu section
gentle quest
well yeah thats kind of how it works
hazy saddle
my brain farted
gentle quest
just like any other variable in asm
gentle quest
np
gentle quest
why people need per cpu section
storing cpu local state
wanton grove
storing cpu local state
just allocating it with allocator would be enough ?
gentle quest
just allocating it with allocator would be enough ?
☠️
how the hell do u know which cpu ur on
wanton grove
im storing cpu num to struct in gs
gentle quest
yeah so you use percpu
just emulated and slow af
this stores all of the data in gs
not just the idx
so any operation on cpu local is 1 cycle and atomic
hazy saddle
this also means i wont need the CPU local head thing mangarm does iirc
because i can refer to gs: cleanly
gentle quest
yep
hazy saddle
i think for the BSP i can always reuse the orignal section too?
as long as i dont have loadable modules that need cpu_local
gentle quest
technically u can i guess
neon karma
the workflow is
```c
DEFINE_PER_CPU(current_thread);
```
in asm:
```x86asm
mov ...
do you happen to know if a similar thing be done on aarch64?
hazy saddle
im pretty sure this is some fuckery of CISC arches
gentle quest
do you happen to know if a similar thing be done on aarch64?
well one cycle locals are not possible on aarch64 no
neon karma
yea, because then this is not an option for me
gentle quest
why not
neon karma
I'd like to not have two cpu local implementations for different arches
gentle quest
why not?
if one arch allows the implementation to be much more optimized
cpu locals in general are very arch specific
hazy saddle
since im only on the BSP rn this is all i need? and then later i calcuate the delta
gentle quest
e.g. on linux they store the current thread pointer in a scratch reg as you've seen
yeah
dont forget to load 0 to gs base
hazy saddle
yes yes. IIRC my gdt should do that for me?
anything else you think i should add to the _Generic's?
hazy saddle
oh thats pretty neat
though _Generic can do it comptime iirc?
ig if anything i can be explicit where its annoying
gentle quest
a fixed switch on sizeof is definitely also comptime
hazy saddle
gentle quest
yes yes. IIRC my gdt should do that for me?
only if u explicitly load something into the gs
that causes a base reload
hazy saddle
ahh
floral kettle
the trick is what i load into GS_BASE is not actually a gs base, rather it's an ...
this is very nice, I hadn't though about using the gsbase like that. I were also thinking that I could do this on msvc and without opts it kinda works but it really fucks it up when you enable optimizations lol (and even on clang the code is questionable) ```c
#include <intrin.h>
char a = 20;
char load(void) {
return __readgsbyte((unsigned long long) &a);
}
gentle quest
why is it so bad on msvc?
so msvc just straight up generates incorrect code?
and i guess clang just does the right thing but not rip relative
at least the load is still atomic
floral kettle
yeah it looks like msvc got rid of the gs segment and in the clang code there is a redundant eax eax move + the lea isn't merged to the load
coarse oasis
the mov eax, eax truncates to 32 bits no?
gentle quest
maybe 'a' is actually annotated with gs somewhere?
no way they have a bug that bad there
coarse oasis
the mov eax, eax truncates to 32 bits no?
or am i tweaking
floral kettle
the mov eax, eax truncates to 32 bits no?
ah yeah, I forgot that the intrinsics take 32-bit offsets and that also makes this approach not work at all with those 💀
gentle quest
well your kernel shouldnt be larger than 4gb right
coarse oasis
sure but you lose potential sign bits
coarse oasis
gs would need the upper 32 bits of the address + lower 32 bits of the offset ig
hazy saddle
https://godbolt.org/z/67MW6511o
it looks fine for me?
gentle quest
we're talking about msvc targets
rotund furnace
this is very nice, I hadn't though about using the gsbase like that. I were also...
well well well
if it isn't msvc being stupid
gentle quest
yeah this cant be real
hazy saddle
msvc is not a c compiler, its a source code mutilator that sometimes manages to emit valid x86 and aarch64 code
urban vale
nobody asked but the preferred way to do a cpu-local structure on mips was to map it into the last page of the address space using a manually created tlb entry and then access it using negative offsets from the zero register
rotund furnace
that's fucking horrifc
hazy saddle
sounds good for RISC tbh
urban vale
so youd put it at the same virtual address on every cpu but it would be mapped to a different phys addr
gentle quest
same virtual on x86 because virtual on x86 is before segmentation is applied
then its linear
floral kettle
yeah this cant be real
I think it is, without O2 it generates a proper gs relative access and if you use O2 + put the address to a volatile variable then it also works (though the code is very much unoptimized)
gentle quest
u should really submit a bug report to them
hazy saddle
in nasm it should be like?
mov [rel gs:syscall_scratch_space], r15
gentle quest
and it was fixed a month ago
neon karma
I'd need to extern "C" my variables if I did that smh
hazy saddle
i dont think nasm is getting the symbol right
this compiles jist fine
src/syscalls/syscalls.asm:9: warning: 32-bit relative section-crossing relocation [-w+reloc-rel-dword]
src/syscalls/syscalls.asm:10: warning: 32-bit relative section-crossing relocation [-w+reloc-rel-dword]
src/syscalls/syscalls.asm:13: warning: 32-bit relative section-crossing relocation [-w+reloc-rel-dword]
src/syscalls/syscalls.asm:34: warning: 32-bit relative section-crossing relocation [-w+reloc-rel-dword]
src/syscalls/syscalls.asm:53: warning: 32-bit relative section-crossing relocation [-w+reloc-rel-dword]
src/syscalls/syscalls.asm:54: warning: 32-bit relative section-crossing relocation [-w+reloc-rel-dword]
src/syscalls/syscalls.asm:57: warning: 32-bit relative section-crossing relocation [-w+reloc-rel-dword]
do you know why nasm isnt allowing this proper?
global syscall_handler
extern execute_syscall
extern current_thread
extern syscall_scratch_space
syscall_handler:
swapgs
mov [rel gs:syscall_scratch_space], r15
mov r15, [rel gs:execute_syscall]
mov [r15 + 32], rsp
mov rsp, [r15 + 24]
mov r15, [rel gs:syscall_scratch_space]
...
it just fully double faults and accepts bunk externs and it compiles fine. am i gonna have to remake this in GAS? 
gentle quest
u probably just didnt convince it to use the right encoding
hazy saddle
I’m too lazy tbh
Il just be a freak and mix Nasm and GAS
I like NASMs macro syntax too much for my ISRs
but now i have proper fast CPU locals
ty for this method infy
gentle quest
i like gas because i can do #include and pass it my C flags and it works out of the box
hazy saddle
gentle quest
also it autoamtically defines __ASSEMBLER__ i can even share header files
between c and assembly
hazy saddle
Oh sick
rotund furnace
ew :P
floral kettle
you don't need the q postfixes etc here btw
hazy saddle
sick
rotund furnace
so?
hazy saddle
nothing ig
now im gonna make my idle thread set the preemption flag
so when idleing any IRQ wakes the system up
rotund furnace
I just yield :P
hazy saddle
well then it wont sleep :p
rotund furnace
but that's prob a better idea
hazy saddle
Now CPU local works
But like good
And its atomic
And nothing used fixed blocks in assembly etc
floral kettle
the reason it didn't work on nasm is probably because you accidentally used execute_syscall instead of current_thread
gentle quest
lmao
urban vale
Now CPU local works
okay now do node-local
hazy saddle
okay now do node-local
I’d possibly store inside CPU local a pointer to node local
But I don’t even have SMP right now
I’m not ready for NUMA shit
hazy saddle
I’d possibly store inside CPU local a pointer to node local
I could even use fs maybe
But that may be annoying
hazy saddle
Ya know
I wonder why astral dosnt do this?
And why it uses fucking CR2
Like how much of a rework is that under astral?
To just use a fucking thing like that
Similarly to Linux and mine
neon karma
And why it uses fucking CR2
// TODO
hazy saddle
Oh yeah timers are now separated from preemption
The timers just set the CPUlocal flag
round shoal
kalajadu
wanton grove
And why it uses fucking CR2
free register ig
hazy saddle
What?
wanton grove
if cr2 is rw why not to use it
hazy saddle
free register ig
It’s a m serializing write
wanton grove
a
hazy saddle
Which makes it slow as balls
And it also causes a vmexit
Which makes it even more slow as balls
hazy saddle
kalajadu
This is what linux does BTW
It just puts it into the unused part of the TSS
hazy saddle
.
round shoal
bro really translated "kalajadu" to "black majick"
hazy saddle
daemon is a background process
floral kettle
And it also causes a vmexit
why would it cause a vmexit?
it wouldn't really make any sense given that there are no control bits or anything in there and even for regs that do like cr3 they don't always cause vmexits (it depends on what you configure the vmx/svm settings to and what's supported by the cpu)
neon karma
I thought of the same thing lol
hazy saddle
why would it cause a vmexit?
Or was that cr8 that did that?
neon karma
I believe so
floral kettle
its possible that it does yeah, though again it depends on how you configure it as + what's supported
full fern
Like how much of a rework is that under astral?
Not really that much, I just got bigger fish to catch
I really really want to be the first one to run steam for shits and giggles
gentle quest
I really really want to be the first one to run steam for shits and giggles
is that still blocked on net memes
full fern
is that still blocked on net memes
Networking on wine is weird and some things work and other don't BUT
I have been consistently getting it to crash in a new way which makes me assume it detected an update
Before it would just keep trying to check and inevitably give a "pls connect to the internet" error after 10-20 seconds
(now that only happens when I forget to run netd, which I really should improve on at some point to run automatically in /etc/rc)
hazy saddle
Not really that much, I just got bigger fish to catch
Understandable
hazy saddle
Updated todo
meson & jinxcpu localunite prremption from schedular (which is pretty much already the case but i dont wana push my half assed messy patches)- make generic linked list helpers & support circular ones too
- make the chunked linked list helper funcs ontop
- implement a generic bitmap allocator for things
- implement the VFS
Now it’s just linked lists and then bitmaps and the VFS
hazy saddle
i think ive got atomic INC/DEC for cpu local now
this works under -O0 and generates the proper opcodes etc
yeah this should work
though 64 bit writes may need to increment and decremt the preemption counter using the INC/DEC operations?
but thoes remain single instruction
and it means they wont cause issues™
Though I wonder what other causes may cause the compiler to do that?
Maybe having a set of them that handle the preemption counter themselves before doing the load/store
To be used inside of places where preemption can happen
Though this depends on what actually is using CPU local and how much of an issue it is if it ends up cross core?
Though in the cases where I don’t wana be preempted I’d possibly need that to last longer than a basic load/store
Eg if I’m doing an operation where I need to fetch the LAPIC ID and use it for somthing to map an IRQ to me
hazy saddle
though 64 bit writes may need to increment and decremt the preemption counter us...
So this ends up back to being fine
neon karma
yeah I have inc/dec for the preemption counter and I manually disable it if I need to when I'm reading a cpu local
or a few helpers that do it, if I just want to load/store
hazy saddle
yeah I have inc/dec for the preemption counter and I manually disable it if I ne...
So I should be able to for x86 use a CPU local INC/DEC for the preemption counter and be all good™ and then have some helpers to do it for that
Though for plain load and stores (that generate one instruction) I think it’s fine for setting a Boolean etc
Depends on the section but it will work then
hazy saddle
Apparently my kernel has enough for fork
Which means the VFS is delayed again babyyyy
I’m going to rework process spawning
So that way I can pass data to new threads
And then I am going to implement fork
I have everything I need for CoW too
Which means I can get bash
Without any VFS syuff
It would be pretty useless
But I am going to at least get it loading
And in theory it should work
And I will have fork and exec()
wanton grove
Which means the VFS is delayed again babyyyy
stop procrastinating your vfs
hazy saddle
I’m thinking with fork() I can have a special lock for it?
Syscalls will take it when doing anything to do with virtual memory etc
But many can take it at once as long as the fork() call or anything else similarly critical needs it etc
That way fork() can have exclusive access to do its work when needed
But other things can just go wild when not fork()ing and use the locks of the backing data structures to keep it safe?
Yeah I would need two locks I think
One lock for the data structure it’s self that’s locked when modifying it
And one lock to prevent those functions from being called when they shouldn’t
wanton grove
Syscalls will take it when doing anything to do with virtual memory etc
do you have lock in vmm
hazy saddle
Eg in a fork()
hazy saddle
do you have lock in vmm
Yes
But I need a lock I can take more long term for these data structures
hazy saddle
I am curious what Linux does here?
If only it were free and open source
And I could just go look for myself
Ehh like with what JW said I shouldn’t prematurely optimize this
Having a regular process local mem lock isn’t horrible
And then I still have the local locks for the underlying things if I change it up etc
clear bison
Get it working first
hazy saddle
Yeah exactly
I have everything I need
Though I should make my process creation a bit more flexible first anyways
hazy saddle
Though I should make my process creation a bit more flexible first anyways
I’m gonna change it so it runs a fixed assembly stub that gets the actual RIP to jmp to and a pointer to a structure that gets passed as an argument
That way I can actually deliver data to it
And I need a stub to conform to the ABI
sterile plank
I’m thinking with fork() I can have a special lock for it?
i recommend Mul-T-Lock Cronus High Security Deadbolt
very good lock
neon karma
I recommend a self-sealing stem bolt
hazy saddle
I recommend a self-sealing stem bolt
What about a latchkey lock
neon karma
I can give you hundred gross self-sealing stem bolts
I don't know what else to do with them
hazy saddle
How did you get so many self-sealing stem bolts
neon karma
I traded five thousand wrappages of cardassian yamok sauce for them
hazy saddle
😭
rotund furnace
I traded five thousand wrappages of cardassian yamok sauce for them
and how did you get five thousand wrappages of cardassian yamok sauce
neon karma
some dumbass waiter in my uncle's bar accidentally ordered them so I took it off their hands
hazy saddle
I now properly handle passing arguments to threads
I’m now gonna make a way to spawn threads generically from a path to an elf file
Yeah not really a huge thing
And I should have supported this for ages
neon karma
kernel shell 
wanton grove
when font change
hazy saddle
It dosnt even do any shell like things
Also I’m literally working on this so I can nuke this thing from the sky
No more “shell” in the kernel
hazy saddle
kernel shell <:nooo:750466870993223750>
By the end of this week I am having bash ported
And if I don’t I will eat a raw onion
I have the stuff I need for fork and everything
neon karma
big bite
neon karma
you don't need fork to run bash
hazy saddle
Yeah but to make it run apps properly
neon karma
apps 😭
full fern
By the end of this week I am having bash ported
Chop chop
hazy saddle
This I fail il record another video like this
hazy saddle
me when i used to not zero registers when jumping to userspace
leaking kernel data
sudden mirage
me when i used to not zero registers when jumping to userspace
does zeroeing it affect the user program?
hazy saddle
does zeroeing it affect the user program?
it zeros the registers on entry to userspace. arguments go onto the stack instead
and then the libc manages that
sudden mirage
and then the libc manages that
ah ok
hazy saddle
i should be able to just poop out a fork now
i think
I should have everything
And now I have a proper way to create processes with arguments
So I can have a fork() trampoline thread
Which handles things differently than making a new process
Because it has to some restore registers
Which may be a bit cursed because il have a
mov rsp, [rsp]
And il give the assembly function a pointer to use as the stack
hazy saddle
ah ok
But yeah if you don’t do this the user app can technically be getting some kernel data leaked
Also make sure to always use the 32bit register versions
It’s faster
sudden mirage
makes sense
wanton grove
32 bit is for virgins when real chad use 128bit
sudden mirage
32 bit is for virgins when real chad use 128bit
actual chads use 512 bit registers
true oak
actual chads use 512 bit registers
actual chads have a supercomputer and use 4096 bit registers
sudden mirage
actual chads have a supercomputer and use 4096 bit registers
damn didnt think of that
hazy saddle
32 bit is for virgins when real chad use 128bit
on x86-64 zeroing a register with xor takes up less code size with the 32bit versions and it still clears the entire full register
And some CPUs don’t see xor r64 r64 as dep breaking
I think with my new build system
Bash shouldn’t be horrid
And I can just get inspired by emmas jinx recipie and patch
Bash like
Won’t do much
But I can atleast have it “ported”
IIRC I should have enough sysdeps for it to load?
hazy saddle
genuily
what the fuck
why did updating mlibc cause everything to break
every app in diffrent ways
hello world?
just fine
doom? GPFs in the middle of some enviroment shit
badapple? jumps to zero
:/
hazy saddle
why does everything make it WORSE?????
mlibc wont even compile
after updating to master
hazy saddle
?????
warped ermine
frigg.cpp
hazy saddle
Kernel Shell> hello
Started playing HELLO in userspace
Kernel Shell> Entering ld.so
ldso: Executable PHDRs are at 0x10000040
rtld: Loading (executable)
rtld: TLS of (executable) mapped to 0x-120, size: 288, alignment: 8
rtld: tcb allocated at 0x20101220, size = 0x90
rtld: tls allocated at 0x201010c0, size = 0x160
rtld: wrote tls image at 0x20101100, size = 0x120
rtld: Running DT_PREINIT_ARRAY functions
rtld: Initialize (executable)
rtld: Running DT_INIT function
rtld: Running DT_INIT_ARRAY functions
In function doInitialize, file ../options/rtld/generic/linker.cpp:1318
__ensure((object->initArraySize % sizeof(InitFuncPtr)) == 0) failed
!!! mlibc panic !!!
Kernel Shell> badapple
Started playing BAD APPLE in userspace
Kernel Shell> Entering ld.so
ldso: Executable PHDRs are at 0x10000040
rtld: Loading (executable)
rtld: TLS of (executable) mapped to 0x-120, size: 288, alignment: 8
rtld: tcb allocated at 0x2bff60, size = 0x90
rtld: tls allocated at 0x2bfe00, size = 0x160
rtld: wrote tls image at 0x2bfe40, size = 0x120
rtld: Running DT_PREINIT_ARRAY functions
rtld: Initialize (executable)
rtld: Running DT_INIT function
rtld: Running DT_INIT_ARRAY functions
rtld: Object initialization complete
Leaving ld.so, jump to 0x100009dc
qemu: terminating on signal 2
why does THIS one fail
?????
I call bullshit on the universe
and doom only prints the panic
and i know i did a clean build
A full entire clean rebuild
Doom still only panics
And prints no extra debit debug shit
What the actual fuck
hazy saddle
“Mlibc sucks I’m porting musl”
- speaking from a crazed madwomen going insane from fixing this
hazy saddle
claude fixed it
cc: @rotund furnace
200k fucking tokens
😭
it was just told "fix it"
it fucking did it
my elf loader was cooked
soft snow
my elf loader was cooked
huh
hazy saddle
cc: <@798918348360515614>
Thanks for using your Claude tokens water
💀
That shit was funny
Watching it go for like an hour
Just thinking on loop
soft snow
claude fixed it
you got Claude subscription
soft snow
Thanks for using your Claude tokens water
🙏
hazy saddle
he just fed my kernel into there
soft snow
<:PSP1G_pspTrollar:1434611394782691429>
Lol
soft snow
Why not
soft snow
I use Codex for debugging sometimes
hazy saddle
it eventualy shat out like 20 lines of fixes
soft snow
Its half baked tbh
soft snow
after 200k tokens
Damn you should test codez
hazy saddle
it even tried to find bugs inside mlibv
soft snow
It's fast asf
soft snow
it even tried to find bugs inside mlibv
It always does
soft snow
Codex hallucinates so much though
It's sad.
I only use it if I'm physically looking at the files
rancid viper
only reason I'd really use codex over claude is cuz its cheaper
rancid viper
yeah
if @rotund furnace just did "fix elf loader" woul've been done in a minute
but no
point it at the entire fucking distro
impressive it found the issue at all
hazy saddle
it was fucking funny
hazy saddle
point it at the entire fucking distro
And the entire fucking libc
I do gotta say
The new mlibc sysdep format is good
soft snow
only reason I'd really use codex over claude is cuz its cheaper
*free
rancid viper
eh
money for amount of querying u can do is cheaper
sure the free plan exist if that is enough
rancid viper
im not a student
viral bison
also, copilot is like 10 buck a month else
hazy saddle
https://git.evalyngoemer.com/evalynOS/mlibc-evalynos/commit/5e7489a44a9889ec52981e649ceb7df1d5a9e46c
force push
The new mlibc version is running well
Time to uhhh
Port bash
Il get the shit for it later to be full full
But I’m not eating an onion
also by popular demand
make bootstrap now also makes the initramfs
and make initramfs regenerates it
wanton grove
The new mlibc version is running well
still no vfs
hazy saddle
still no vfs
soft snow
money for amount of querying u can do is cheaper
how much does claude even cost
hazy saddle
bash works
ts is fucking cursed tho
yes let me just sed some random shit out
cold barn
btw for clarity I would recommend using [.] instead of \.
also while [[:space:]] can be nice if you want it to match unicode whitespace characters, or at least theoretically depending on what the implementation supports, you can also use \s. err, you might need to use -E with that, I'm not sure. but \s is part of many common regex implementations and should be at least "all ascii whitespace characters" if not "all whitespace characters in my character set"
cold barn
btw for clarity I would recommend using `[.]` instead of `\.`
same for $, basically anything except [/], I guess? reducing the amount of \ in a regex statement is almost always nice, especially when there's special meaning for it like in sed, where if it's in the s//replace/ part, it places a previous capture group there
neon karma
c++ reflection so good that they added it in sed
cold barn
I go by the rule of thumb that if something need the literal value, and I can't just do -F (for grep), then it should use [] instead of \ except where it explicitly doesn't make sense
hazy saddle
also while `[[:space:]]` can be nice if you want it to match unicode whitespace ...
it seems \s makes things break
but the ither things worked
sed -i 's/\bgetenv[.]o\b//g' lib/sh/Makefile
sed -i 's/\bstrchrnul[.]o\b//g' lib/sh/Makefile
sed -i 's/[$]{LIBOBJDIR}strchrnul[$]U[.]o[[:space:]]*//' lib/sh/Makefile
why not to use patches instead of sed
cold barn
it seems \s makes things break
odd, just use a named character classes (not the "official" name but I can't find one) then I guess lol
cold barn
😔
oh uh
it is probably smart enough to realize you don't mean a range or exact match here, but you should [] the { and }
though, those might only be special characters on extended/perl regex
hazy saddle
With jinx it is a pinned Debian thing running this so technically it shouldn’t ever cause issues
cold barn
ah ok
hazy saddle
The rest of the bash sysdeps shouldn’t be terrible tho?
Atleast I can stub them for the most part
Very heavily
hazy saddle
💀
tf is bash trying to do
🥀
addr2line -e extras/initramfs/bash.elf 0x00000000100003ca
/build_dir/builds/mlibc/build/../subprojects/frigg/include/frg/manual_box.hpp:20 (discriminator 2)
cr7 did not do that error AEEE
hazy saddle
what?
round shoal
tf is bash trying to do
steal some cash in the name of GNU
wanton grove
elf loader is broken
wanton grove
then elf loading is baddd
hazy saddle
ahh
my get time fucntion was garbgo
int Sysdeps<ClockGet>::operator()(int clock, time_t *secs, long *nanos) {
*secs = 0;
*nanos = 0;
}
for somre reason
this isnt valid
neon karma
```c
int Sysdeps<ClockGet>::operator()(int clock, time_t *secs, long *nanos)...
I'm also permanently stuck in 1970
hazy saddle
I'm also permanently stuck in 1970
It’s chill here
Also me when I forget a return value
😭
clear bison
Also me when I forget a return value
Wouldn't the compiler or LSP complain about that
neon karma
Also me when I forget a return value
no warnings?
neon karma
God the new sysdep system is nice
what else is different in it?
hazy saddle
what else is different in it?
it uses operators and a list of sysdeps you give
so it can do compile time asserts better
and if it cant find a sysdep it can now fallback to ENOSYS
yoooo its bash printing
neon karma
nice
neon karma
sadly i think il need my VFS to go further
for getcwd?
neon karma
stupid quick reaction row
neon karma
nuh uh
hazy saddle
what even is the sysdep for this
neon karma
probably getcwd :P
neon karma
i cant find it in others sysdeps
hazy saddle
there we go
Let’s see what happens with more memory
I know my shit is so fucking inefficient esp with any file ops
round shoal
invalid opcode??????
your fat32 or filesystem is broken,happened to me yesterday
hazy saddle
no
i dont have any filesystems
:p
bash really be suckling my memory
💀
That was with 8gb ram
round shoal
i dont have any filesystems
what the fuck are you doing without any filesystem
hazy saddle
what the fuck are you doing without any filesystem
running doom and bash
round shoal
u have ahci or ata
hazy saddle
no
round shoal
just start with ata
hazy saddle
no
round shoal
id rater just do nvme and virti
virtio
normal turret
hazy saddle
Now onto Trisha takanaw for the forecast for today
hazy saddle
okay
the only issue
is faking stdin enough
That’s a bash port
The input layer is all sorts of FUCKED
But it works
marble ridge
neat!
true oak
That’s a bash port
the onion was spared
marble ridge
would be nice if i could read anything but still
hazy saddle
Once I get my VFS for a proper input layer
It should be usable
Then I can tape on fork()
hazy saddle
~~would be nice if i could read anything but still~~
Let me change the font for this
marble ridge
nah don't bother, i trust you lol
i can just zoom in the image
it's just basically unreadable on my 1440p screen without zooming in though
true oak
so you restart the shell instead of waitpid?
true oak
why does it print all of this stuff after echo
hazy saddle
why does it print all of this stuff after echo
errors
soft snow
your returning works atleast :D
hazy saddle
ya know what
im not fixing that issue
until i get my VFS done
Somehow bash internally is just fucking up?
It’s getting the data it copied to it?
But it keeps stale data in its buffer
Making it fuck up bad
neon karma
I'm guessing you're doing a unix like vfs?
hazy saddle
I do want to yes
hazy saddle
wtf
bash-5.3$ echo "helloooo"
helloooo
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$ echo "testtt"
testtt
o
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$
HOME
also setpgid
hazy saddle
i did stub that last one
int Sysdeps<SetPgid>::operator()(pid_t pid, pid_t pgid) {
return 0;
}
bash-5.3$ echo "hello"
0x65 0x63 0x68 0x6f 0x20 0x22 0x68 0x65 0x6c 0x6c 0x6f 0x22 0xa
hello
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$ echo "a"
0x65 0x63 0x68 0x6f 0x20 0x22 0x61 0x22 0xa
a
llo
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$ echo "a"
0x65 0x63 0x68 0x6f 0x20 0x22 0x61 0x22 0xa
a
llo
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$ echo "hello"
0x65 0x63 0x68 0x6f 0x20 0x22 0x68 0x65 0x6c 0x6c 0x6f 0x22 0xa
hello
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$
i am sending the right bytes?
It looks like bash isn’t clearing its buffer
it seems when i send a \n here
it breaks the buffer?
That’s a \n being printed there
And the rest of the buffer still has data
So bash gets angry?
viral bison
maybe bc you are breaking?
hazy saddle
wdym?
hazy saddle
maybe not?
viral bison
idk, i can't see the full code
viral bison
XD
hazy saddle
bash-5.3$ stdin: count=1024 rpos=0 len=0
echo "test"
test
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$ stdin: count=1024 rpos=12 len=12
echo "t"
t
st
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$ stdin: count=1024 rpos=9 len=9
well, now to this bug
hazy saddle
its getting a \n stuck somwhere i think
viral bison
this should be before any other case btw
else there might be cases where pressing return or any other chars will do a buffer overflow
hazy saddle
echo "test"
test
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$ stdin: count=1024 rpos=12 len=12
echo "1"
1
st
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$ stdin: count=1024 rpos=9 len=9
bash-5.3$ ot set terminal process group (1): Operation not implemented (ENOSYS)
bash: no job control in this shell
bash: //.bashrc: Access denied (EACCESS)
bash-5.3$ stdin: count=1024 rpos=1 len=1
yeah well, stdin is not the problem here i think
hazy saddle
ah
hazy saddle
It should be mostly standard stdin?
viral bison
but only print it
hazy saddle
maybe try to not put \n inside the buffer?
I’d I do that it never gets the /n for bash to run the command
Let me try again
yeah
it ends up thinking its an exit
viral bison
ah
viral bison
i think bash wants one
hazy saddle
Idk why it says ENOSYS tho
viral bison
lol
viral bison
But why would it do *this*
maybe this can help?
i think the first error is the problem
anyway, it should still work
it will just throw the warnings / errors at you
hazy saddle
what even are the sysdeps for pgrp
viral bison
wdym pgrp?
hazy saddle
viral bison
ah that
hazy saddle
i dont get any mlibc errors
wait
In function mlibc::sysdep_return_t<Tag, Args ...> mlibc::sysdep_or_enosys(Args&& ...) [with Tag = Tcgetattr; Args = {int&, termios*&}; sysdep_return_t<Tag, Args ...> = int], file ../options/internal/include/mlibc/all-sysdeps.hpp:53
???
right there
viral bison
coarse oasis
return 0 where
viral bison
go look here for the chat log
hazy saddle
Me when UB is undefined
neon karma
```c++
int Sysdeps<SetPgid>::operator()(pid_t pid, pid_t pgid) {
ret...
process groups are for job control
hazy saddle
Yeah I’m just stubbing the fuck out of it
neon karma
real
viral bison
well, go look at that chatlog
coarse oasis
return 0 where
ftr this should have emitted a warning
viral bison
maybe the compiler auto added return 0?
hazy saddle
wtf is this
coarse oasis
no the compiler only does that for main in hosted environments
hazy saddle
Could it be anything to do with this
💀
I swear it feels like something inside of bash isn’t being cleared out on \n
What else is strange
Is every newline
It complains about no job control
true oak
It complains about no job control
thats why i thought you were just restarting the shell every time
hazy saddle
I’m not
😭
Like
Wtf is it doing
Perplexity AI: bash still can run commands but something is fucked up with its buffer when reading stdin. It is not these things https://www.perplexity.ai/search/bash-still-can-run-commands-bu-buPycj8ERzy2PlKTcjFtyQ
@viral bison it says its an issue with the tty
viral bison
lol
hazy saddle
@viral bison
https://www.perplexity.ai/search/kernel-shell-hello-icanon-on-e-4mPZ46uGTZW0gY1DUhnzJA#3
this is what it says now
viral bison
this is what it says now
did it fix it?
viral bison
you should not use perplexity for code
hazy saddle
Yeah it fucking sucks ass
hazy saddle
viral bison
bc it can actually get the context
viral bison
install copilot
hazy saddle
This can be applied to the dev branch rn
viral bison
i'll try
viral bison
If it dosnt il regenerate
hazy saddle
Let me regenerate it
Just a moment
@viral bison
do a git pull
then apply that
and use this for the libc
viral bison
codeberg
hazy saddle
Yeah codeberg is always a bit behind
Use the git server
@viral bison did ze patch work?
viral bison
yeah, now it applied
hazy saddle
lol
There we go
Now it should just be a make bootstrap
Then make run
And then BASH
And you should be able to reproduce it
If you put the zip file in the right spot
viral bison
ye, making now
hazy saddle
How’s the new build system btw?
viral bison
@hazy saddle
viral bison
ah
viral bison
How’s the new build system btw?
wait, it has to build its own GCC?
hazy saddle
Yes
viral bison
seems like so, pain
viral bison
how much time do i need for it? XD
viral bison
i5 12th gen (mobile)
viral bison
viral bison
XD
hazy saddle
im zipping the jinx folder
so make initramfs? should just work after
and or cd ./jinx && ./jinx rebuild mlibc
viral bison
so make initramfs? should just work after
and if i run make bootstrap?
hazy saddle
hm?
viral bison
says it cannot find limine.h
hazy saddle
uhh
cd ./jinx && ./jinx host-rebuild limine edk2-bin
and where did you put the folder
viral bison
still nope
viral bison
hazy saddle
viral bison
this normal after running bash?
viral bison
it's adding stuff here
hazy saddle
cd ./jinx && ./jinx rebuild mlibc && ./jinx rebuild bash && cd .. && make initramfs && make run && cd ./jinx
that will rebuild mlibc & bash btw
viral bison
k thx